2. What is Colour It?
● The project is about automatically colourizing gray-scaled images
● Colourization of black and white images will be done without any assistance from humans
● Our system will be mapping a sufficient amount of statistical dependencies between an image’s
semantics and textures with a coloured version of it to make it look as real as possible.
3. An example of the working of our system on an input image (right)
4. How?
The users will upload a grayscale image onto the application and the software will provide them with a
plausible results by producing vibrant and realistic colourization of the images
5. How?
● The system consists of an application that is implemented as a feed forward pass in a computational neural network at test time
and is trained over a million colour images
● A self supervised learning technique is explored that uses raw image data as a source of supervision
6. Goals
and
Objectives
● Give any user the capability to process their grayscale images
into coloured adaptations with minimum processing time
● The objective of the image being the closest adoption of the
ground truth can be further extended to medical image
processing
● It can be further implemented in video editing bringing a
plethora of black and white films to life
● To provide a plausible output image that puts human
cognitive skills to test
● Gain command over convolutional neural networks
7. Motivation
and
Background
● Image colourization has been the topic of debate amongst the
public since decades
● The artistic appeal to it can be recognized in the simple act of
bridging the gap between the generations by reproducing
memories that are more relatable to the current time
● Previously, colour was manually added which seemed very
tedious and time consuming
● Combination of machine learning and image processing equip
even the most novice of users to achieve this without any
hassle.
8. Constraints
● Time taken for processing an image might vary from subject to
subject
● Input image can only be of a 256 x 256 size.
● Some images with distorted pixels might not yield correct
results after the processing
● An unfamiliar image that was not present in the training
dataset might hinder the accuracy of predicted colour space
● To achieve maximum accuracy, the training sets need to be
extensive and rigorous. However, obtaining a dataset like this is
not only difficult but also, average sizes vary in TBs which
makes it extremely time consuming to obtain in our internet
conditions